How Our Intelligence Models Work

What It Measures

The vulnerability score (0-100) estimates how likely an incumbent contractor is to be displaced when their contract is re-competed. Higher scores indicate weaker incumbent positions and better opportunities for challengers.

Scoring Factors

The score is computed from 6 weighted factors, each derived from publicly observable contract characteristics:

Score Interpretation

Data Source

Computed nightly from 23M+ contract awards in USAspending.gov (FY2022-2026). Bridge contract detection uses keyword matching on contract modification descriptions (terms: "bridge", "interim", "extension", "continuation", "temporary").

Known Limitations

• CPARS performance ratings are not publicly available and cannot be factored in
• Bridge detection relies on description keywords and may miss some bridges or flag false positives
• The score does not account for political factors, incumbent relationships, or non-public information
• Scores are relative within our dataset — a score of 40 does not mean a 40% chance of displacement

What It Measures

When a federal contract ends and the work is picked up by a different contractor, that's usually called “displacement.” But the raw rate bundles together very different things — some of those losses weren't competitive at all. Displacement Analytics segments every matched re-compete into five outcome buckets so you can see what the incumbent actually faced.

Classification Buckets

How We Match Old → New Pairs

For every expired contract above $1M in FY22-FY25, we find the most likely follow-on by searching for contracts that share:

• Same agency, NAICS code, and PSC code
• Action date within 60 days before to 120 days after the expired contract's end date
• Obligated amount between 0.5× and 2× the expired contract's value
• Same place of performance (if specified on the expired contract)
• Different parent IDV (if the expired had one) — to avoid matching task orders to their own parent

When multiple candidates match, we pick the one whose action date is closest to the expired contract's end date. A small fraction of expired contracts don't produce any match (truly terminated work, major restructures, etc.) and are excluded from the dataset.

Vehicle Awardee Detection

We derive vehicle awardee lists directly from USAspending task-order data — every contract with a parent PIID is grouped by that parent, and the union of unique vendor UEIs on those task orders is treated as the vehicle's awardee pool. This covers GWACs, MACs, BOAs, and agency-specific IDVs without requiring us to maintain a curated list of vehicle names.

Current coverage: ~41,800 parent vehicles derived from ~620K task orders. Newer vehicles without sufficient task-order history may be classified as “unknown vehicle” until enough data accumulates.

How To Read The Numbers

Three headline rates for any filtered slice (FY / agency / NAICS):

• Raw Displacement Rate — traditional “% of re-competes where the incumbent lost.” Combines structural and competitive losses.
• Structural Rate — displacements where the incumbent couldn't have competed on the same terms (set-aside rotation + vehicle migration).
• True Competitive Loss Rate — displacements where the incumbent had a seat at the table and still lost. This is the rate that actually predicts competitive dynamics in your market.

Known Limitations

• Matching is probabilistic — we can't know with certainty that contract A is the follow-on to contract B. Tight agency+NAICS+PSC+POP+value filters keep the false-match rate low, but edge cases exist (e.g., work genuinely split across two follow-ons).
• Set-aside classifications use the raw USAspending labels — subtle changes in how an agency codes set-asides year over year may be over- or under-counted as rotation.
• Vehicle awardee detection is based on observed task-order history. A vehicle awardee who has won zero task orders will not appear in our awardee list.
• Excludes contracts below $1M, contracts without a PSC code, contracts ending before FY22 Q2, and contracts where no follow-on can be identified.
• Does not yet segment prime-to-sub flips (where the original incumbent continues to perform the work as a subcontractor on the new award) — that requires subaward data we don't currently ingest at scale.

Model Type

Bayesian-style heuristic model. This is not a trained machine learning model — it applies empirically-weighted factors to a base rate derived from market competition data. Each factor acts as a multiplier on the base probability.

How It Works

1. Base rate is calculated from the average number of competing offers in your NAICS code. If the average is 4 offers, the base rate is 1/4 = 25%.

2. Seven adjustment factors modify the base rate up or down:

Output Range

Capped at 2-85%. We do not output probabilities above 85% or below 2% because no contract outcome is certain.

Accuracy & Calibration

We are actively backtesting this model against historical outcomes. Preliminary results indicate the model is directionally correct — contracts scored >30% win probability have historically been won at higher rates than those scored <15%. Full calibration analysis with confidence intervals will be published here when complete.

Known Limitations

• This is a heuristic model, not a trained ML model — factors and weights are based on domain expertise and published research, not gradient descent
• The model cannot account for proposal quality, key personnel, technical approach, or other non-quantifiable factors
• Accuracy varies by NAICS code — codes with fewer historical awards have less reliable base rates
• The model assumes your company profile inputs are accurate
• Past win rates do not guarantee future outcomes

Methodology

For each NAICS code + agency + set-aside + fiscal year combination, we compute the statistical distribution of historical award values:

• 25th percentile (P25): 75% of awards were above this value
• Median (P50): The midpoint — half above, half below
• 75th percentile (P75): 25% of awards were above this value
• Average: Mean value (often skewed by large outliers)

Recommended Bid Range

We recommend bidding in the 25th-50th percentile range. This range is competitive enough to win on price while high enough to pass price realism reviews. Bids below P25 risk being flagged as unrealistic. Bids above P75 need strong justification.

Data Quality

Benchmarks are computed from pre-aggregated statistics, updated nightly. We require a minimum of 5 awards in a NAICS+FY combination to produce a benchmark. Combinations with fewer awards do not display results.

Known Limitations

• Award values include modifications — some records represent net-zero modifications, not competitive awards
• The dataset does not distinguish between base awards and option year exercises
• FY2022 shows higher median values than FY2023-2026, likely due to data composition differences
• Labor rate information from GSA CALC is separate from award-level pricing

Processing Schedule

• 3:00 AM ET daily: Re-compete candidates flagged and vulnerability scores computed
• 4:00 AM ET daily: Pricing, competitor, and spending statistics recomputed
• 4:30 AM ET daily: Email alerts sent to subscribed users
• 1st of each month: Full bulk data refresh from USAspending.gov